Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

added disc

Shen-Chenhuid8fab103f4eb
diff --git a/configs/vae/train/video_disc.py b/configs/vae/train/video_disc.py
index 0e10d62..9967c5f 100644
--- a/configs/vae/train/video_disc.py
+++ b/configs/vae/train/video_disc.py
@@ -42,13 +42,13 @@ discriminator = dict(
)
# discriminator hyper-parames TODO
-discriminator_factor=1,
+discriminator_factor=1
discriminator_start=-1
-generator_factor=0.5,
-generator_loss_type="hinge",
-discriminator_loss_type="hinge",
-lecam_loss_weight=None,
-gradient_penalty_loss_weight=None,
+generator_factor=0.5
+generator_loss_type="hinge"
+discriminator_loss_type="hinge"
+lecam_loss_weight=None
+gradient_penalty_loss_weight=None
# loss weights
perceptual_loss_weight = 0.1 # use vgg is not None and more than 0
diff --git a/opensora/models/vae/vae.py b/opensora/models/vae/vae.py
index ac320b7..ec2643d 100644
--- a/opensora/models/vae/vae.py
+++ b/opensora/models/vae/vae.py
@@ -153,7 +153,7 @@ class VideoAutoencoderPipeline(nn.Module):
return self.temporal_vae.get_latent_size(self.spatial_vae.get_latent_size(input_size))
def get_temporal_last_layer(self):
- return self.temporal_vae.last_layer[0]
+ return self.temporal_vae.decoder.conv_out.conv.weight
@property
def device(self):
diff --git a/scripts/train-vae.py b/scripts/train-vae.py
index 58e33ed..60400d7 100644
--- a/scripts/train-vae.py
+++ b/scripts/train-vae.py
@@ -277,9 +277,9 @@ def main():
vae_loss += adversarial_loss
# Backward & update
+ optimizer.zero_grad()
booster.backward(loss=vae_loss, optimizer=optimizer)
optimizer.step()
- optimizer.zero_grad()
# Adversarial Discriminator loss
@@ -295,9 +295,9 @@ def main():
)
disc_loss = weighted_d_adversarial_loss
# Backward & update
+ disc_optimizer.zero_grad()
booster.backward(loss=disc_loss, optimizer=disc_optimizer)
disc_optimizer.step()
- disc_optimizer.zero_grad()
all_reduce_mean(disc_loss)
running_disc_loss += disc_loss.item()